Agents that do the work and show it
kern agents run on your machine, use real tools, remember everything, and publish their own dashboards. Not chatbots — autonomous workers that maintain their own UI.
Open source. Self-hosted. Ready in 60 seconds.
The agent intranet
Every function in your company runs an agent. Every agent publishes a dashboard. Put them behind one URL. That's your intranet — except it builds itself.
Sales agent
Pipeline by stage, deal velocity, stalled deals. Click 'follow up' — the agent drafts and sends the email.
Infra agent
Service health, deploy status, cost trends. Red indicator? Click — the agent explains and proposes a fix.
Finance agent
Burn rate, runway, spend breakdown. Cloud cost spike? Click — the agent traces it to a service.
Support agent
Ticket queue, response times, themes. Click a ticket — the agent drafts a reply from your docs.
One brain, every channel
Terminal, browser, Telegram, Slack — one continuous session. The agent knows who's talking, what channel it's in, and what happened 10,000 messages ago.
Terminal ─────┐ Web UI ───────┤ Telegram ─────┤── one session Slack ────────┘
Memory that compounds
Conversations segmented by topic, summarized into a hierarchy, compressed into context. Semantic recall over everything. The agent gets better the longer it runs.
L2 ▪▪ (weeks) L1 ▪▪▪▪▪▪ (days) L0 ▪▪▪▪▪▪▪▪▪▪ (topics) raw ──────────────────
Your infra, your data
Runs on your laptop, server, or homelab. The whole agent is a git-tracked folder. Pay only for API tokens — or use Ollama for fully local, zero-cost inference.
my-agent/ ├── .kern/config.json ├── IDENTITY.md ├── KNOWLEDGE.md ├── knowledge/ └── notes/
See inside the agent's brain
Every conversation is segmented, summarized, and searchable. Five tabs show exactly what the agent knows and how it thinks.
Semantic segments
Conversations split by topic, rolled up into hierarchical summaries. Color-coded by token density.
Recall search
Semantic search over every past conversation. Find decisions, configs, or context from months ago.
Sessions
Message activity, channel breakdown, role distribution. See how the agent spends its time.
Context window
Token budget visualization. System prompt, compressed summaries, raw messages — see what fits.
One brain, every channel
Other frameworks start a fresh session for each channel. Same agent, but it forgets what happened on Telegram when someone asks in Slack. kern feeds every message into one continuous session — nothing is lost.
Typical approach
Context lost between channels. Agent repeats work, forgets decisions, can't connect the dots.
kern
Every message from every interface, one context window. The agent connects what you said on Telegram to what someone asked in Slack.
What ships today
Not a roadmap. Primitives you can use right now.
Agent-built dashboards
Agents create HTML dashboards with live data injection. Rendered in a side panel or inline in chat.
Multi-modal
Images, PDFs, files across every channel. Vision pre-digest, PDF extraction, dedicated analysis tools.
Desktop app
Native macOS via Tauri. Tray icon, Cmd+1-9 agent switching, direct connections.
Prompt caching
Three cache breakpoints. 99% mid-turn hits, 10x cost reduction. Automatic for Anthropic.
React web UI
Flat and bubble layouts, syntax highlighting, infinite scroll, multi-agent sidebar with live status.
Real tools
bash, read, write, edit, grep, webfetch, websearch, pdf, image, render — full system access.
4 providers
OpenRouter, Anthropic, OpenAI, Ollama. Mix models per role — chat, embeddings, summaries, vision.
Heartbeat
Agents wake periodically — review notes, update knowledge, reach out if needed. Autonomous maintenance.
Plugin architecture
Dashboard, media, recall, notes extracted as plugins. Extend with lifecycle hooks, no core changes.
How kern compares
Different tools, different bets on what agents should be.
| kern | Claude Code | Codex | Cowork | OpenClaw | |
|---|---|---|---|---|---|
| Focus | General-purpose | Coding | Coding | Work tasks | General-purpose |
| Agent-built UI | Dashboards | ✗ | ✗ | Artifacts | ✗ |
| Memory | Hierarchical DAG | CLAUDE.md | Per-task | Limited | Resets daily |
| Session model | One unified | Per-project | Per-task | Per-channel | Per-channel |
| Channels | 5 interfaces | Terminal | API | Slack | 20+ platforms |
| Self-hosted | ✓ | ✗ | ✗ | ✗ | ✓ |
| Model choice | Any provider | Claude only | OpenAI only | Claude only | Any provider |
| System access | Full (shell, fs) | Git repos | Sandbox | Sandbox | Full (shell, fs) |
| Open source | MIT | ✗ | ✗ | ✗ | Apache 2.0 |
Your first agent in 60 seconds
Pick a model. Start talking. It remembers everything from here.